DANN (Domain-Adversarial Neural Network) pytorch

Paper: Domain-Adversarial Training of Neural Networks

Prerequisite

• python 3.7 (Anaconda)

• pytorch >= 1.0

• torchvision >= 0.2.2

Dataset

• SVHN(Source dataset), MNIST(Target dataset)

• Download from torchvision

Description

• DANN.ipynb : DANN model and training algorithm

• NN.ipynb : Baseline model and training algorithm to compare with DANN

• The models were trained by 20~30 epochs respectively

• You can see the graphs of loss and accuracy in ipynb

• Implementation and Result

• GRL (Gradient Reversal Layer)
• 3 Modules (Feature extractor, Classifier, Discriminator)
• Data preprocessing (Resize 28x28, gray scale, normalization [-1, 1], )
• 2 Datasets (SVHN, MNIST)
• Test Accuracy (on MNIST test set)

• DANN: 70.64 %
• Baseline model (Source only model): 57.80 %